A data center flywheel energy storage load prediction and scheduling method

Abstract

The application relates to the technical field of data center energy management, and discloses a data center flywheel energy storage load prediction and scheduling method, which comprises the following steps: obtaining data center operation data and flywheel energy storage state data, and integrating the data into a time sequence data matrix; performing time sequence feature decomposition on the matrix to obtain a load dynamic and energy storage response mode feature matrix; performing weight distribution and feature aggregation through a dynamic correlation network and a sliding window mechanism to generate a collaborative feature matrix; and generating a scheduling instruction containing a charging and discharging strategy based on a decision tree model and a historical case library. The method realizes multi-source data fusion analysis, improves prediction accuracy and scheduling efficiency, enhances energy utilization rate and power supply reliability, and is suitable for data center energy management.

Description

Technical Field

[0001] This invention relates to the field of data center energy management technology, specifically a method for predicting and scheduling flywheel energy storage load in data centers. Background Technology

[0002] With the rapid development of information technology, data centers, as the core infrastructure of the digital economy, are experiencing explosive growth in scale and energy consumption. During data center operation, the volatility and uncertainty of power load are significant, such as dynamic load changes in server clusters, power consumption fluctuations in cooling systems due to ambient temperature, and periodic peaks in network traffic. These factors place extremely high demands on the stability of the power grid supply for data centers. Simultaneously, to address the energy crisis and environmental pressures, data centers urgently need to improve their renewable energy absorption capacity and energy utilization efficiency. Flywheel energy storage technology, with its advantages of fast response speed, long cycle life, and strong environmental adaptability, has become an important choice for data center energy storage systems.

[0003] Existing flywheel energy storage systems for data centers have significant shortcomings in load forecasting and scheduling. Traditional methods typically treat load forecasting and energy storage scheduling separately, lacking collaborative analysis of data center operational data and flywheel energy storage status data. This results in an inability to accurately capture the dynamic correlation between load fluctuations and energy storage responses. For example, forecasting based solely on historical power load data, without fully considering the impact of energy storage status parameters such as flywheel speed and bearing temperature on scheduling strategies, can easily lead to unreasonable charging and discharging timing of the energy storage system, affecting power supply reliability and energy utilization.

[0004] At the technical implementation level, existing solutions mostly employ single time-series analysis models or empirical rules for scheduling, making it difficult to effectively handle multi-dimensional and unstructured operational data. For example, traditional models cannot achieve refined feature decomposition and dynamic weight allocation for the coupled effects of multiple parameters such as peak power load, cooling system power consumption, and network traffic fluctuations, resulting in insufficient load forecasting accuracy and a lack of flexibility and adaptability in scheduling strategies. Furthermore, existing methods fail to fully utilize the safety thresholds and historical operating cases of energy storage systems when generating scheduling instructions, potentially leading to problems such as equipment overload or low operating efficiency.

[0005] As data centers evolve towards higher density and greater intelligence, higher demands are placed on load forecasting and scheduling of flywheel energy storage systems. This necessitates deep fusion and collaborative analysis of multi-source data to construct a dynamically responsive load-energy storage characteristic model, thereby improving forecast accuracy and the scientific rigor of scheduling strategies. It also requires the introduction of advanced machine learning algorithms and optimization mechanisms to achieve dynamic adjustment of feature weights and local feature optimization, enhancing the system's adaptability to complex operating scenarios. Furthermore, a scheduling rule generation mechanism based on historical cases and safety thresholds is needed to ensure the safe and efficient operation of the energy storage system. Therefore, researching a method that integrates data center operational data with flywheel energy storage status data to achieve load-energy storage collaborative characteristic analysis and dynamic scheduling has significant theoretical and practical application value. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting and scheduling flywheel energy storage load in data centers, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting and scheduling flywheel energy storage load in a data center, the method comprising:

[0008] Data center operation data collected by power monitoring devices at multiple time points within a predetermined time period, and flywheel energy storage status data collected by energy storage monitoring devices at the same multiple time points;

[0009] Pattern recognition and collaborative analysis are performed on the operational data and energy storage status data of the multiple acquisition time nodes to obtain the load-energy storage collaborative feature matrix;

[0010] Based on the load-energy storage collaborative feature matrix, flywheel energy storage load scheduling instructions are generated;

[0011] Specifically, pattern recognition and collaborative analysis are performed on the operational data and energy storage status data from the multiple data acquisition time points to obtain a load-energy storage collaborative feature matrix, including:

[0012] The operational data from the multiple acquisition time nodes are integrated into an operational parameter time series data matrix according to time order and parameter category, and the energy storage status data from the multiple acquisition time nodes are integrated into an energy storage status time series data matrix according to time order and status index.

[0013] Time series feature decomposition is performed on the operating parameter time series data matrix and the energy storage status time series data matrix to obtain the load dynamic mode feature matrix and the energy storage response mode feature matrix, respectively.

[0014] Dynamic weight allocation and feature aggregation are performed on the load dynamic mode feature matrix and the energy storage response mode feature matrix to obtain the load-energy storage collaborative feature matrix.

[0015] Preferably, the operating data includes peak power load, cooling system power consumption, and network traffic fluctuation parameters.

[0016] Preferably, the energy storage status data includes flywheel speed, bearing temperature curve, and energy storage unit charge / discharge rate.

[0017] Preferably, time series feature decomposition is performed on the operating parameter time series data matrix and the energy storage state time series data matrix to obtain the load dynamic mode feature matrix and the energy storage response mode feature matrix, respectively, including:

[0018] The time series data matrix of the operating parameters is input into the feature extraction module based on the time series decomposition algorithm to extract the combined features of trend components, periodic components and residual components to form the load dynamic mode feature matrix.

[0019] The energy storage state time series data matrix is ​​input into the state feature extraction module based on the pattern recognition encoder to extract multi-dimensional state change features, forming the energy storage response mode feature matrix.

[0020] Preferably, the load-storage coordinated feature matrix is ​​obtained by dynamically weighting and aggregating the load dynamic mode feature matrix and the energy storage response mode feature matrix, including:

[0021] The load dynamic mode feature matrix and the energy storage response mode feature matrix are input into a dynamic correlation network to generate a feature weight allocation matrix.

[0022] The feature weight allocation matrix is ​​locally optimized using a sliding window mechanism to obtain an optimized weight matrix.

[0023] The optimized weight matrix is ​​applied to the load dynamic mode feature matrix and the energy storage response mode feature matrix respectively, and then the features are concatenated and encoded to generate the load-energy storage coordinated feature matrix.

[0024] Preferably, the load dynamic mode feature matrix and the energy storage response mode feature matrix are input into a dynamic correlation network to generate a feature weight allocation matrix, including:

[0025] Calculate the correlation metric between the statistical distribution index of the load dynamic mode feature matrix and the state change index of the energy storage response mode feature matrix;

[0026] A two-dimensional weight distribution map is constructed based on the correlation metric, and the feature weight allocation matrix is ​​generated using an interpolation algorithm.

[0027] Preferably, the feature weight allocation matrix is ​​obtained by performing local feature optimization processing on the sliding window mechanism, including:

[0028] A sliding window is set on the feature weight allocation matrix to perform local feature extraction and obtain the window feature vector;

[0029] After performing principal component dimensionality reduction on the window feature vector, a weighted average operation is performed to reconstruct the optimized weight matrix.

[0030] Preferably, based on the load-energy storage collaborative feature matrix, a flywheel energy storage load scheduling instruction is generated, including:

[0031] The load-energy storage collaborative feature matrix is ​​input into the scheduling rule generation module based on the decision tree model to output a set of energy storage scheduling parameters;

[0032] The flywheel energy storage load scheduling instruction, which includes a charging and discharging strategy, is generated based on the set of energy storage scheduling parameters.

[0033] Preferably, the load-energy storage coordinated feature matrix is ​​input into the scheduling rule generation module based on the decision tree model to output a set of energy storage scheduling parameters, including:

[0034] The load-energy storage synergy feature matrix is ​​sorted by feature importance to determine key decision dimensions;

[0035] Branching conditions are applied to the key decision dimensions to derive the set of energy storage scheduling parameters.

[0036] Preferably, generating the flywheel energy storage load scheduling instruction containing a charging and discharging strategy based on the energy storage scheduling parameter set includes:

[0037] The set of energy storage scheduling parameters is matched with the historical operation case library to determine the baseline scheduling parameters;

[0038] The baseline scheduling parameters are dynamically adjusted by calculating parameter offsets to generate a charge / discharge strategy range that includes a safety threshold.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] By collecting data center operation data and flywheel energy storage status data through power monitoring and energy storage monitoring devices, the data encompasses multi-dimensional information such as peak power load, cooling system power consumption, network traffic fluctuation parameters, flywheel speed, bearing temperature curves, and energy storage unit charge / discharge rates. This overcomes the limitations of traditional single-dimensional data analysis and enables dynamic correlation modeling between load fluctuations and energy storage responses. By integrating operation data and energy storage status data into time-series data matrices and performing time-series feature decomposition, the system can effectively extract dynamic load pattern features such as trend components, periodic components, and residual components, as well as multi-dimensional state change characteristics, providing rich feature representations for load forecasting and scheduling.

[0041] In terms of feature aggregation and weight allocation, a dynamic correlation network is used to calculate the correlation metric between load and energy storage features, construct a two-dimensional weight distribution map, and generate a feature weight allocation matrix. A sliding window mechanism is then used for local feature optimization, achieving dynamic adjustment and adaptive allocation of feature weights. This mechanism can automatically adjust the relative importance of load and energy storage features according to changes in real-time operating scenarios. For example, it automatically increases the weight of dynamic load features during peak power load periods and enhances the focus on energy storage status features during energy storage system maintenance cycles, thereby improving the feature model's adaptability to complex operating conditions and its prediction accuracy.

[0042] In the process of generating scheduling instructions based on the load-energy storage collaborative feature matrix, a decision tree model is used to rank the importance of features, determine key decision dimensions, and apply branch judgment conditions. Combined with a historical operating case library, pattern matching and dynamic parameter adjustment are performed to generate a charging and discharging strategy range that includes safety thresholds. This process combines data-driven analysis with a rule engine, avoiding the blind reliance on historical data while overcoming the limitations of empirical rules. It can dynamically generate scientific and reasonable scheduling strategies based on real-time features, ensuring that the flywheel energy storage system accurately matches the data center load demand in terms of charging and discharging timing and power control.

[0043] The application of this method offers numerous technical benefits: at the load forecasting level, multi-dimensional data fusion and feature decomposition effectively improve the accuracy of load fluctuation forecasting, reducing power supply instability caused by forecasting errors; at the energy storage scheduling level, dynamic weight allocation and local feature optimization mechanisms enable scheduling strategies to respond in real-time to changes in operating status, improving the charging and discharging efficiency and energy utilization of flywheel energy storage systems; regarding system reliability, the scheduling rule generation mechanism based on safety thresholds and historical cases reduces equipment overload risks and extends the service life of flywheel energy storage units; at the energy management level, this method provides technical support for data centers to integrate renewable energy and participate in grid peak shaving, contributing to the achievement of green and low-carbon development goals for data centers. Furthermore, through modular design and algorithm integration, this method possesses good scalability and compatibility, adaptable to flywheel energy storage systems of different sizes and types, and has broad engineering application prospects. Attached Figure Description

[0044] Figure 1 This is a schematic diagram illustrating the working principle of the data center flywheel energy storage load prediction and scheduling method described in this invention.

[0045] Figure 2 A flowchart for dynamic weight allocation and feature aggregation;

[0046] Figure 3 The flowchart for optimizing the generation of the weight matrix. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Please see Figures 1-3 The present invention relates to a method for predicting and scheduling flywheel energy storage load in data centers, the specific implementation steps of which are as follows:

[0049] Step 1 involves acquiring data center operation data from multiple data collection points within a predetermined time period, collected by the power monitoring device, and flywheel energy storage status data from the same multiple data collection points, collected by the energy storage monitoring device. The predetermined time period can be set according to actual needs, such as daily, weekly, or monthly. The power monitoring device and the energy storage monitoring device collect the corresponding data in real time to ensure the timeliness and accuracy of the data.

[0050] Step 2 involves performing pattern recognition and collaborative analysis on the operational data and energy storage status data from the multiple data acquisition time nodes to obtain a load-energy storage collaborative feature matrix. This step specifically includes the following sub-steps: First, the operational data from the multiple data acquisition time nodes are integrated into an operational parameter time-series data matrix according to time order and parameter categories; then, the energy storage status data from the multiple data acquisition time nodes are integrated into an energy storage status time-series data matrix according to time order and status indicators. Next, time-series feature decomposition is performed on the operational parameter time-series data matrix and the energy storage status time-series data matrix to obtain a load dynamic pattern feature matrix and an energy storage response pattern feature matrix. Finally, dynamic weight allocation and feature aggregation are performed on the load dynamic pattern feature matrix and the energy storage response pattern feature matrix to obtain the load-energy storage collaborative feature matrix.

[0051] Step 3 is executed, and a flywheel energy storage load scheduling instruction is generated based on the load-energy storage collaborative feature matrix.

[0052] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0053] Example 1:

[0054] In this embodiment, the operational data specifically includes peak power load, cooling system power consumption, and network traffic fluctuation parameters. The peak power load is collected in real-time by a power monitoring device at multiple data collection points within a predetermined time period. The collection frequency can be set to once per minute or every five minutes, depending on the data center's operational characteristics, to capture peak power consumption at different time points. This parameter reflects the extreme power consumption of the data center in an instant or short period, such as a sudden surge in power load during peak periods of server cluster processing batch tasks or user access. The collected data is arranged chronologically to form a corresponding data sequence, for example, using timestamps as indexes to record the peak power load value at each data collection time point.

[0055] Cooling system power consumption data is also collected synchronously by power monitoring devices. As a key component in maintaining a constant temperature environment in the data center, the power consumption of the cooling system is closely related to the computing load of the servers. When the server workload increases, more heat is generated, and the cooling system needs to increase its operating power to enhance heat dissipation. Therefore, cooling system power consumption parameters can indirectly reflect the actual business load of the data center. During the collection process, it is necessary to distinguish the power consumption data of different components of the cooling system (such as air conditioning units, fans, etc.) and classify and integrate them according to time nodes and component categories. For example, first arrange the total cooling power consumption of each node in chronological order, and then break it down to the power consumption value of each component to form multi-dimensional cooling system power consumption time-series data.

[0056] Network traffic fluctuation parameters are acquired through network monitoring equipment in the data center and are used to characterize the changes in data transmission volume between the data center and external networks over time. The level of network traffic is directly related to the type of services the data center handles and the volume of user access. For example, video streaming services typically experience higher network traffic during evening hours, while enterprise office data centers experience more concentrated traffic during weekdays. This parameter can be further subdivided into upload and download traffic, recording the traffic values ​​at each collection point and integrating them chronologically into a time-series data sequence of network traffic fluctuations for subsequent analysis of the correlation between network services and power load.

[0057] When integrating operational data, the three types of parameters mentioned above must be structured according to time sequence and parameter category, as required in step 2. Specifically, an operational parameter time-series data matrix should be established, with rows corresponding to data collection time nodes and columns corresponding to different parameter categories (peak power load, cooling system power consumption, and network traffic fluctuation parameters). For example, if the predetermined time period is one day and the data collection time node is once per minute, the matrix contains 1440 rows (24 hours × 60 minutes) and 3 columns (corresponding to the three types of parameters). The data in each column of each row are the peak power load of the data center, the total power consumption of the cooling system, and the network traffic fluctuation value (such as the sum of upload and download traffic) for that minute. During the data integration process, abnormal data needs to be filtered and repaired, for example, by using numerical interpolation of adjacent time nodes or statistical averaging to handle missing or abrupt data, ensuring that the data in the matrix can truly reflect the actual operating status of the data center.

[0058] By accurately collecting and systematically integrating these three types of specific operational data, a time-series data matrix of operational parameters with time-series characteristics and multi-parameter dimensions was formed. This provides structured input data for subsequent steps of time-series feature decomposition of operational data. This matrix not only preserves the dynamic information of each parameter changing over time, but also, through the classification of parameter categories, enables in-depth analysis of different types of load characteristics. For example, it allows for the separate analysis of the correlation between the trend changes of peak power load and the periodic fluctuations of cooling system power consumption, or the exploration of the synchronicity between network traffic fluctuations and peak power load, thus laying the foundation for constructing a load dynamic pattern feature matrix.

[0059] Example 2:

[0060] In this embodiment, the energy storage status data specifically includes flywheel speed, bearing temperature curve, and energy storage unit charge / discharge rate. Among these, flywheel speed is the core parameter reflecting the operating status of the flywheel energy storage device, and it is acquired in real time by a high-precision sensor. This sensor is typically mounted on the flywheel's rotating shaft and uses electromagnetic induction or optical measurement principles to accurately measure the flywheel's rotational speed at each acquisition time point. The acquisition frequency is set according to the characteristics of the flywheel energy storage system, generally at the millisecond level, to capture subtle changes in flywheel speed. Within a predetermined time period, the flywheel speed value at each acquisition time point is recorded chronologically, forming a time-series data of the flywheel speed. For example, within a complete charge / discharge cycle, the flywheel speed gradually increases from its initial value to the rated speed, and then gradually decreases during discharge. By continuously recording these speed values, the energy storage and release process of the flywheel energy storage device can be fully presented.

[0061] Obtaining the bearing temperature profile relies on a network of temperature sensors distributed throughout the flywheel bearing. These sensors, employing thermocouple or thermistor technology, monitor the bearing's temperature changes in real time during operation. Since bearing temperature changes relatively slowly, the acquisition frequency can be set to the second or minute level. During acquisition, not only is the absolute temperature value recorded at each time point, but also the rate of temperature change over time, forming a two-dimensional temperature-time curve. For example, during flywheel startup, the bearing temperature gradually rises, and once it reaches a stable operating state, the temperature remains within a relatively stable range; however, when abnormal wear or insufficient lubrication occurs, the temperature shows a continuous upward trend. Analysis of these temperature curves allows for the timely detection of potential bearing failures.

[0062] The charge / discharge rate of the energy storage unit reflects the energy exchange efficiency between the flywheel energy storage device and the data center power system, and is collected in real time by a power monitoring module. This module, installed in the power conversion unit of the flywheel energy storage system, accurately measures the current and voltage values ​​during the charge / discharge process and calculates the charge / discharge power using the power calculation formula (power = voltage × current). Dividing the charge / discharge power by the rated power of the energy storage unit yields the charge / discharge rate. During the data acquisition process, the charge / discharge rate values ​​at each acquisition time point are recorded in chronological order, forming a time-series data of the charge / discharge rate. For example, during periods of low power load in the data center, the flywheel energy storage device is charging, and the charge / discharge rate is positive; while during periods of high power load, the flywheel energy storage device is discharging, and the charge / discharge rate is negative. By analyzing the variation pattern of the charge / discharge rate, the responsiveness of the flywheel energy storage device to data center power fluctuations can be assessed.

[0063] When integrating energy storage status data, following the requirements of step 2, the three types of parameters mentioned above are integrated according to time sequence and status indicators. First, an energy storage status time-series data matrix is ​​established. The rows of the matrix correspond to the acquisition time nodes, and the columns correspond to different status indicators (flywheel speed, key parameters of the bearing temperature curve, charge / discharge rate). For the bearing temperature curve, its key parameters (such as peak temperature, maximum and average temperature change rate, etc.) are extracted as column elements of the matrix. For example, if the predetermined time period is one hour and the acquisition frequency is once per minute, the matrix contains 60 rows and 3 columns (corresponding to flywheel speed, key parameters of the bearing temperature curve, and charge / discharge rate). During the data integration process, the raw acquired data is preprocessed, including data cleaning, filtering, and standardization. Data cleaning mainly removes abnormal data points caused by sensor failures or communication interference; filtering uses moving average or Kalman filtering algorithms to smooth the data curve and reduce the impact of random noise; standardization converts parameters of different dimensions to a unified numerical range for subsequent feature analysis.

[0064] By accurately collecting and systematically integrating these three types of specific energy storage state data, a time-series data matrix of energy storage state with time series characteristics and multi-parameter dimensions was formed. This matrix not only retains the dynamic information of each parameter changing over time, but also, through the division of state indicators, enables in-depth analysis of different types of energy storage characteristics in subsequent analyses. For example, it allows for the separate analysis of the correlation between the changing trend of flywheel speed and charge / discharge rate, or the exploration of the impact of abnormal fluctuations in bearing temperature curves on the performance of energy storage devices, thus laying the foundation for constructing an energy storage response mode characteristic matrix.

[0065] Example 3:

[0066] In this embodiment, the process of performing time series feature decomposition on the operating parameter time series data matrix and the energy storage state time series data matrix is ​​as follows:

[0067] The time-series data matrix of operating parameters is input to a feature extraction module based on a time-series decomposition algorithm. This module employs decomposition algorithms suitable for multi-dimensional time-series data, such as seasonal decomposition based on statistical models or wavelet decomposition based on signal processing. Taking seasonal decomposition as an example, each column of parameters in the matrix (such as peak power load, cooling system power consumption, and network traffic fluctuation parameters) is first decomposed. During the decomposition process, the time-series data of each parameter is broken down into trend components, periodic components, and residual components.

[0068] Trend components characterize the long-term changing trend of parameters over time, such as the overall upward or downward trend of peak power load in data centers during weekdays, or the continuous increase in cooling system power consumption due to server hardware upgrades. Periodic components reflect the periodic changing patterns of parameters, such as the daily cycle of network traffic fluctuations, which show low traffic at fixed times of the day (e.g., off-peak hours at night) and peak traffic during daytime office hours, or the seasonal periodic changes in cooling system power consumption with the changing seasons. Residual components are the random fluctuations remaining after removing trend and periodic components, typically caused by accidental factors such as sudden business requests or momentary equipment failures.

[0069] By decomposing the time series of each parameter, the combined features of these three components are extracted. For example, for the peak power load sequence, the decomposition yields the slope of the trend component (e.g., an increase of 0.5 kW per hour), the amplitude of the periodic component (e.g., peak fluctuation range within a daily period of ±20%), and the standard deviation of the residual component (e.g., random fluctuation amplitude of ±5 kW). These feature values ​​are arranged according to time nodes and parameter categories to form a load dynamic pattern feature matrix. The rows of this matrix still correspond to the original data acquisition time nodes, while the columns correspond to the trend, periodic, and residual feature dimensions of each parameter after decomposition, thus transforming the original operating data into more physically meaningful dynamic pattern features.

[0070] The energy storage state time series data matrix is ​​input to a state feature extraction module based on a pattern recognition encoder. This module utilizes encoder structures in deep learning (such as Convolutional Neural Networks (CNNs) or Long Short-Term Memory Networks (LSTMs)) combined with pattern recognition algorithms to extract features from the energy storage state time series data. Taking a CNN encoder as an example, the energy storage state time series data matrix (such as the time series of flywheel speed, bearing temperature, and charge / discharge rates) is first converted into a two-dimensional matrix form as input. Then, local and global features are extracted from the data through multi-layer convolutional kernels.

[0071] In flywheel speed feature extraction, the encoder captures dynamic features such as acceleration during the speed increase phase, speed fluctuation rate during stable operation, and speed decrease rate during discharge. For key parameters of the bearing temperature curve (such as peak temperature and temperature change rate), the encoder identifies features such as the starting time point of abnormal temperature rise and the slope of temperature gradient change. For charge and discharge rates, the encoder extracts features such as charging power threshold, discharge power abrupt change point, and charge / discharge switching frequency. These features are mapped into low-dimensional feature vectors through multi-layer nonlinear transformation of the encoder, forming an energy storage response mode feature matrix. The rows of this matrix correspond to the original acquisition time nodes, and the columns correspond to multi-dimensional state change features such as flywheel speed change features, bearing temperature response features, and charge / discharge rate regulation features.

[0072] During the feature decomposition process, two types of matrices require preprocessing. For the time-series data matrix of operating parameters, normalization is first performed to convert parameters with different dimensions (such as kW, kW, Mbps) into dimensionless values, avoiding the impact of dimensional differences on the decomposition algorithm. For the time-series data matrix of energy storage status, a sliding window technique is used to divide the raw data into frames, with each window containing data from N consecutive acquisition time nodes (e.g., N=10 minutes), facilitating the encoder to capture state change patterns within a short period. In addition, both modules are equipped with outlier detection mechanisms, using statistical methods (such as the Z-score method) to identify and correct abnormal feature values ​​that may occur during the decomposition process, ensuring the reliability of the feature matrix.

[0073] Through the aforementioned time-series feature decomposition process, the time-series data matrix of operating parameters is transformed into a load dynamic mode feature matrix, and the time-series data matrix of energy storage status is transformed into an energy storage response mode feature matrix. These two matrices no longer rely on the specific physical dimensions of the original data, but instead describe the operating patterns of data center load and flywheel energy storage devices using abstract feature dimensions. This provides a structured feature representation for subsequent dynamic weight allocation and feature aggregation, enabling collaborative analysis from the perspective of the dynamic correlation between load and energy storage.

[0074] Example 4:

[0075] In this embodiment, the process of dynamically weighting and aggregating the load dynamic mode feature matrix and the energy storage response mode feature matrix is ​​as follows:

[0076] The load dynamic mode feature matrix and the energy storage response mode feature matrix are input into a dynamic correlation network. This network consists of a correlation coefficient calculation layer, a weight distribution map generation layer, and a weight matrix interpolation layer. In the correlation coefficient calculation layer, statistical distribution indicators (such as mean, variance, skewness, and kurtosis) of the load dynamic mode feature matrix and state change indicators (such as rate of change, gradient, and extreme points) of the energy storage response mode feature matrix are calculated pairwise. For example, the Pearson correlation coefficient between the daily variation trend of peak power load and the response delay of flywheel speed is calculated, or the correlation between the fluctuation variance of cooling system power consumption and the bearing temperature change rate is calculated. These calculation results form a two-dimensional correlation coefficient matrix, where each element represents the correlation strength between a load feature and an energy storage feature.

[0077] Based on the correlation coefficient matrix, a two-dimensional weight distribution map generation layer constructs a weight distribution map. This map uses the load feature dimension as the horizontal axis and the energy storage feature dimension as the vertical axis, with the color or height of each coordinate point representing the correlation coefficient value of the corresponding feature pair. Through threshold filtering (e.g., removing weakly correlated pairs with an absolute correlation coefficient less than 0.3) and Gaussian smoothing, a continuous weight distribution surface is generated. The weight matrix interpolation layer employs a bicubic interpolation algorithm to generate a dense feature weight allocation matrix based on the discrete correlation coefficient matrix, accurately distributing weight values ​​to each load-energy storage feature pair.

[0078] A sliding window mechanism is used to perform local feature optimization on the feature weight allocation matrix. A sliding window of size M×N is set (e.g., M=5 load features, N=3 energy storage features), sliding across the feature weight allocation matrix with a step size S (e.g., S=1). For each window position, eigenvectors are extracted and a local covariance matrix is ​​constructed. Principal component analysis (PCA) is used to perform eigenvalue decomposition on the covariance matrix, retaining principal components whose contribution rate exceeds a set threshold (e.g., 85%), thus achieving local feature dimensionality reduction. For the dimensionality-reduced eigenvectors, time-decay weights are assigned based on their temporal position in the original data (e.g., the weight of nearest neighbor time points is 1, and the weight of distant time points decays exponentially), and then a weighted average is calculated to obtain the locally optimized eigenvectors. The optimization results of all windows are reorganized according to their original positions to form an optimized weight matrix.

[0079] During the feature aggregation phase, the optimized weight matrix is ​​applied to both the load dynamic mode feature matrix and the energy storage response mode feature matrix. Specifically, feature weighting is achieved through matrix multiplication. For example, for the load dynamic mode feature matrix A (dimension T×P, where T is the time point and P is the number of load features) and the optimized weight matrix W (dimension P×Q, where Q is the number of aggregated features), the weighted matrix A' = A×W is calculated. Similarly, the same operation is performed on the energy storage response mode feature matrix to obtain the weighted matrix B'.

[0080] Feature concatenation encoding is performed on A' and B'. A feature concatenation approach is used, concatenating the two matrices column-wise to form a joint feature matrix of dimension T×(2Q). To enhance the interaction between features, a secondary weighting is applied to the joint feature matrix using an attention mechanism. The attention mechanism calculates the importance score for each feature dimension; features with higher scores are assigned greater weight in subsequent processing. Specifically, a multilayer perceptron (MLP) is used to perform a non-linear transformation on the joint feature matrix to generate an attention weight vector. This vector is then multiplied element-wise by the joint feature matrix to achieve adaptive feature enhancement.

[0081] Finally, the attention-weighted joint feature matrix is ​​subjected to dimensionality reduction using Linear Discriminant Analysis (LDA). This reduces the feature dimension to K dimensions (e.g., K=Q) while maintaining class separability, resulting in the final load-storage collaborative feature matrix. This matrix retains the dynamic correlation information between load and energy storage features while reducing redundancy through weight optimization and feature aggregation, providing more representative and discriminative input features for subsequent generation of scheduling instructions based on the decision tree model.

[0082] An adaptive learning mechanism was implemented throughout the dynamic weight allocation and feature aggregation process. By comparing the changing trends of feature weights across different time windows, the sliding window size M, N, and step size S were dynamically adjusted. Based on the sparsity and correlation distribution of the feature matrix, an appropriate interpolation algorithm and dimensionality reduction threshold were automatically selected. Furthermore, a regularization term was introduced to prevent overfitting. An L1 / L2 regularization penalty term was added to the optimization objective function to constrain the parameter size of the weight matrix and ensure the model's generalization ability.

[0083] Example 5:

[0084] In this embodiment, the process of generating flywheel energy storage load scheduling instructions based on the load-energy storage collaborative feature matrix is ​​as follows:

[0085] The load-energy storage synergy feature matrix is ​​input into the scheduling rule generation module based on a decision tree model. This module uses the CART (Classification and Regression Tree) algorithm to construct the decision tree model. During the model training phase, based on historical load and energy storage status data and corresponding scheduling parameter labeled samples, the splitting conditions of branch nodes are determined by minimizing Gini impurity or mean square error. In the feature importance ranking stage, the decision tree model automatically calculates the feature importance score based on the information gain contribution of each feature in the branching process. For example, the trend component of peak power load, the response delay of flywheel speed, and the periodic component of cooling system power consumption may be identified as key decision dimensions, and their importance scores are significantly higher than other features.

[0086] When applying branch judgment conditions on key decision dimensions, the decision tree recursively divides the value range of each key feature, starting from the root node. For example, if the peak trend characteristic value of the power load in the load dynamic mode is greater than a set threshold (e.g., 1.2 times the average peak value of the previous 30 minutes), and the flywheel speed in the energy storage response mode is lower than the safety lower limit (e.g., 60% of the rated speed), then the charging branch is triggered; if the peak trend characteristic value of the power load is lower than the threshold and the flywheel speed is higher than 80% of the rated speed, then the discharging branch is triggered. Through multi-level condition judgments, the corresponding set of energy storage scheduling parameters is finally output at the leaf nodes, including the preliminary value range of parameters such as charging and discharging power, duration, and switching time.

[0087] After obtaining the set of energy storage scheduling parameters, pattern matching needs to be performed with a historical operation case database to determine the baseline scheduling parameters. The historical operation case database stores typical operation scenarios and their corresponding successful scheduling schemes over a past period (e.g., the previous 12 months). Each case includes a load-energy storage feature vector, scheduling parameters, and execution result feedback. Pattern matching uses a cosine similarity algorithm to calculate the similarity between the current energy storage scheduling parameter set feature vector and the feature vectors in the case database one by one, selecting the top N cases with the highest similarity (e.g., N=5) as the matching results. The baseline scheduling parameters are determined by taking the weighted average of the scheduling parameters of these N cases. The weight is proportional to the similarity score; for example, the case with the highest similarity has a weight of 0.4, the second highest is 0.3, and so on.

[0088] When dynamically adjusting baseline scheduling parameters using parameter offset calculations, the differences between the current operating state and historical cases are first analyzed. For example, if the peak temperature of the current bearing temperature curve is 5°C higher than the historical average, the charging and discharging power needs to be reduced accordingly to avoid bearing overheating; if the real-time value of the network traffic fluctuation parameter exceeds the fluctuation range of historical cases, the charging and discharging duration is extended to cope with possible load surges. The parameter offset calculation is based on quantitative indicators of the differences, such as temperature difference and flow difference, and converts the differences into adjustment values ​​for scheduling parameters using linear transformation formulas (e.g., offset = difference value × adjustment coefficient). The adjustment coefficient is preset according to the safety operation specifications of the energy storage device; for example, for every 1°C increase in temperature, the charging power is reduced by 0.5% of the rated power.

[0089] When generating a charge / discharge strategy range that includes safety thresholds, hard safety thresholds are first determined, such as the minimum safe flywheel speed (to avoid stalling), the maximum alarm value for bearing temperature (to prevent mechanical failure), and the upper and lower limits of charge / discharge power (to protect power electronic devices). These thresholds are preset based on the physical characteristics of the energy storage device and the manufacturer's technical specifications, and the scheduling parameters must not exceed these ranges under any circumstances. Then, combined with dynamically adjusted baseline scheduling parameters, a soft strategy range is generated with the hard safety thresholds as boundaries. For example, if the baseline charging power is 80% of the rated power and the hard safety upper limit is 90%, the charging power strategy range is set to [70%, 85%], which both reserves margin for real-time adjustment and ensures that the safety limits are not reached.

[0090] In the refinement of the strategy range, fuzzy logic rules are introduced to handle the coupled effects of multiple parameters. For example, when the peak trend of the power load shows that the load is continuously rising and the flywheel speed is close to the minimum safe value, the fuzzy logic controller outputs a priority adjustment factor for charging power based on two input variables: "load rise rate" and "speed margin". This dynamically compresses the upper boundary of the strategy range to avoid insufficient subsequent discharge capacity due to overcharging. The strategy range is ultimately presented in the form of a parameter list, including information such as the charging and discharging power range (e.g., charging power: 300-500kW, discharging power: -500 to -300kW), time window (e.g., charging period: 00:00-06:00), and safety restrictions (e.g., suspending charging and discharging when bearing temperature > 75℃), forming a complete flywheel energy storage load scheduling instruction.

[0091] Throughout the generation process, a manual intervention interface was implemented. When the system detects abnormal characteristics (such as rare feature combinations in the load-energy storage coordination feature matrix) or when a safety threshold is about to be breached, it automatically triggers an alert and pauses command generation, awaiting confirmation from operations and maintenance personnel or manual parameter adjustment. Furthermore, a scheduling command execution feedback mechanism is established to record the actual execution effect of each command (such as charging and discharging efficiency, load fluctuation smoothness), and the historical operation case library is regularly updated, deleting failed cases and adding new typical cases to ensure that the accuracy of pattern matching gradually improves over time.

[0092] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data center flywheel energy storage load forecasting and scheduling method, characterized in that, The method comprises the following steps: acquiring data center operation data collected by a power monitoring device at multiple collection time nodes within a predetermined time period, and flywheel energy storage state data collected by an energy storage monitoring device at the multiple collection time nodes; performing pattern recognition and collaborative analysis on the operation data at the multiple collection time nodes and the energy storage state data at the multiple collection time nodes to obtain a load-energy storage collaborative feature matrix; generating a flywheel energy storage load scheduling instruction based on the load-energy storage collaborative feature matrix; wherein the pattern recognition and collaborative analysis on the operation data at the multiple collection time nodes and the energy storage state data at the multiple collection time nodes to obtain a load-energy storage collaborative feature matrix comprises: integrating the operation data at the multiple collection time nodes into an operation parameter time sequence data matrix according to time sequence and parameter category, and integrating the energy storage state data at the multiple collection time nodes into an energy storage state time sequence data matrix according to time sequence and state index; respectively performing time sequence feature decomposition on the operation parameter time sequence data matrix and the energy storage state time sequence data matrix to obtain a load dynamic mode feature matrix and an energy storage response mode feature matrix; performing dynamic weight distribution and feature aggregation on the load dynamic mode feature matrix and the energy storage response mode feature matrix to obtain the load-energy storage collaborative feature matrix; generating a flywheel energy storage load scheduling instruction based on the load-energy storage collaborative feature matrix comprises: inputting the load-energy storage collaborative feature matrix into a scheduling rule generation module based on a decision tree model to output an energy storage scheduling parameter set; generating the flywheel energy storage load scheduling instruction containing a charge-discharge strategy according to the energy storage scheduling parameter set.

2. The data center flywheel energy storage load prediction and dispatch method of claim 1, wherein, The operation data includes power load peak, cooling system power consumption, and network traffic fluctuation parameters.

3. The data center flywheel energy storage load prediction and dispatch method of claim 2, wherein, The energy storage state data includes flywheel speed, bearing temperature curve, and energy storage unit charge-discharge rate.

4. The data center flywheel energy storage load prediction and dispatch method of claim 3, wherein, respectively performing time sequence feature decomposition on the operation parameter time sequence data matrix and the energy storage state time sequence data matrix to obtain a load dynamic mode feature matrix and an energy storage response mode feature matrix comprises: inputting the operation parameter time sequence data matrix into a feature extraction module based on a time sequence decomposition algorithm to extract combined features of trend components, periodic components, and residual components, forming the load dynamic mode feature matrix; inputting the energy storage state time sequence data matrix into a state feature extraction module based on a pattern recognition encoder to extract multi-dimensional state change features, forming the energy storage response mode feature matrix.

5. The data center flywheel energy storage load prediction and dispatch method of claim 4, wherein, performing dynamic weight distribution and feature aggregation on the load dynamic mode feature matrix and the energy storage response mode feature matrix to obtain the load-energy storage collaborative feature matrix comprises: inputting the load dynamic mode feature matrix and the energy storage response mode feature matrix into a dynamic correlation network to generate a feature weight distribution matrix; performing local feature optimization processing on the feature weight distribution matrix through a sliding window mechanism to obtain an optimized weight matrix; The optimized weight matrix is respectively applied to the load dynamic mode feature matrix and the energy storage response mode feature matrix, and then feature splicing coding is performed to generate the load-energy storage collaborative feature matrix.

6. The data center flywheel energy storage load prediction and dispatch method of claim 5, wherein, The load dynamic mode feature matrix and the energy storage response mode feature matrix are input into a dynamic correlation network to generate a feature weight distribution matrix, including: calculating the correlation measure value between the statistical distribution index of the load dynamic mode feature matrix and the state change index of the energy storage response mode feature matrix; constructing a two-dimensional weight distribution map based on the correlation measure value and generating the feature weight distribution matrix through an interpolation algorithm.

7. The data center flywheel energy storage load prediction and dispatch method of claim 6, wherein, The feature weight distribution matrix is locally optimized by a sliding window mechanism to obtain an optimized weight matrix, including: setting a sliding window on the feature weight distribution matrix to extract local features to obtain a window feature vector; performing principal component dimension reduction processing on the window feature vector and then performing a weighted average operation to reconstruct the optimized weight matrix.

8. The data center flywheel energy storage load prediction and dispatch method of claim 1, wherein, The load-energy storage collaborative feature matrix is input into a scheduling rule generation module based on a decision tree model to output a set of energy storage scheduling parameters, including: performing feature importance sorting on the load-energy storage collaborative feature matrix to determine the key decision dimension; applying a branch judgment condition on the key decision dimension to derive the set of energy storage scheduling parameters.

9. The data center flywheel energy storage load prediction and dispatch method of claim 8, wherein, According to the set of energy storage scheduling parameters, the flywheel energy storage load scheduling instruction containing the charging and discharging strategy is generated, including: performing pattern matching on the set of energy storage scheduling parameters and the historical operation case library to determine the reference scheduling parameter; performing dynamic adjustment on the reference scheduling parameter through parameter offset calculation to generate a charging and discharging strategy interval containing a safety threshold.

Images